As a result 1,568 birds were counted. The accuracy over- and underestimation was determined using visual counts against a
defined section with 500 bird objects. During the visual quality check of the 500 bird objects 477 gulls
95.4 could actually be verified. Thereby 65 objects were identified incorrectly as a bird and at the same time 42 birds
were not classified as such. When counting by hand several main characteristic errors of the recognition of bird objects
could be identified: incorrectly recognized objects were mainly pieces of wood, stones and locally bright spots in green spaces.
In addition, dark and or birds in the shade of bushes were not detected. Birds too close together classified as one object. Also
some birds in blurred image regions were not classified correct- ly. Main reasons for image blur was on the one hand the wide-
angle lens of the Sony NEX5 and on the other hand the stitching software that generated a blurred area around the seamlines of
adjacent images.
3. Automatic Bird recognition and Bird count 2012
Building on the experience of 2011, in 2012 the study area could be extended by using the MD4-1000, which has a flight
time of about 25 - 30 minutes per battery. Also the flight alti- tude above ground was slightly increased to 55 m. An aerial
survey of the relevant parts of the island could be carried out successfully on 25.5.2012 recording a total of 629 images. The
external conditions were perfect, almost windless and cloudless sky. The aerial survey of the island was carried out by four
individual flights to temporarily change the battery and to min- imize the required time of flight by a clever choice of the start-
ing points. To ensure that the camera does not stop to take images during the flight because of lack of electricity, the cam-
era battery was changed between flights. Accompanying the aerial survey some signalized control points
were measured with a RTK GPS at one edge of the island. Since the birds were in the business of breeding, island could not be
accessed for control point measurements. Thus, the ground control points mainly served to fit the block in terms of position
and scale. The photogrammetric data analysis was performed using the
software Pix4D, which was used as a web service. The results of the photogrammetric data processing were a digital ortho
photo, a digital surface model DSM and a 3D-point cloud. According to the report the position and height accuracy 1
σ at the five control points was approximately 5 mm.
3.1 Automatic bird recognition and bird count 2012
The method of automatic bird count in 2012 is basically similar to the approach for 2011. Only the size of the filter to eliminate
small misallocations and large objects were changed slightly. The “bird census” for 2012 showed a total of 1,945 birds, see
Figure 3. The classification quality was validated with 605 objects. There were 560 items verified as birds. On closer in-
spection 45 objects turned out to not be a bird, but as something else. At the same time 31 birds were not recognized as such.
That means 97.6 of the birds were detected. The increase in the classification accuracy in 2012, can be clearly attributed to
the better image quality. Figure 3. Identified bird objects red dots from UAS aerial
survey of 25.5.2012 on the birds reserve island Langenwerder 3.2
Identification of breeding birds
For the bird census counting all individuals is an important factor. The number of clutches is of more interest to the orni-
thologists, than just the number of birds, because the average number of eggs per clutch is known and the reproduction poten-
tial is an important number. Since the size of the nest is not visible from the air - after all the birds sit on the nest, it is of
great interest to determine at least the number of breeding pairs. The conducted aerial survey during the morning hours was
basically at an optimal time, as in most breeding pairs, one partner at this time is out foraging for food. So the number of
counted birds approximately matches with the breeding pairs. Visually the breeding birds can be distinguished from the stand-
ing birds. A closer inspection reveals that the standing birds produce a small shadow, as shown nicely in Figs 4a and 4c.
The direction and length of the shadow is determined by the position of the sun, resulting from the geographical position
54° 01 N, 11° 29 E and the time about 10:00 clock EST, as well as the size of the upright standing birds. Thus, the shadow
falls from an elevation of about 40° toward 290° WNW. With a bird size of about 30 to max. 40 cm, the length of the shadow
is 34 - 46 cm.
This contribution has been peer-reviewed. 171
The dark shadow area can be easily classified and spectrally separated from the rest of the image. Around the birds a sectoral
buffer was defined with a diameter of 40 cm. The shadow mask was overlaid with the bird buffer zones 4,137 objects. It turns
out that many wrong assignments have been eliminated in further processing steps. The errors in this initial step are due to
the fact, that the vegetation grassland itself introduces many small shadow elements and furthermore birds like to breed in
tall grass, which creates the shadows. A first area filter removes many small shadow elements that can occur next to breeding
birds. Another filter eliminates shadow objects that are not adjacent to a bird, which means less than 15 cm from the bird
centroid. A final filtering step selects those birds that are not in the tall grass 15 cm. The grass height is determined from the
point cloud see section 6.1. The dark shadow area can be easily classified and spectrally
separated from the rest of the image. Around the birds a sectoral buffer was defined with a diameter of 40 cm. The shadow mask
was overlaid with the bird buffer zones 4,137 objects. It turns out that many wrong assignments could be eliminated in further
processing steps. The errors in this initial step are due to the fact, that the vegetation grassland itself introduces many small
shadow elements and furthermore birds like to breed in tall grass, which creates the shadows. A first area filter removes
many small shadow elements that can occur next to breeding birds. Another filter eliminates shadow objects that are not
adjacent to a bird, which means less than 15 cm from the bird centroid. A final filtering step selects those birds that are not in
the tall grass 15 cm. The grass height is determined from the point cloud see section 6.1.
Finally, 54 birds could be identified that do not breed, but pur- sue other activities. A visual inspection revealed that the strate-
gy was not really successful. Rather, the success rate was only about 74. Among others, one reason is birds often breed in the
immediate vicinity to some higher grasses which cause shad- ows. At the same time several standing birds were also not
recorded properly, e.g. because they stand with their longitudi- nal axis in east-west direction, so they only cast very small
shadows, see figure 4a-c. As a conclusion perhaps 50 - 100 birds are not in the business of breeding, so that the number of
breeding pairs on Langenwerder in 2012 is about 1,850 - 1,900.
3.3 Accuracy assessment
